6
$\begingroup$

I want to compute a fast matrix-vector product using a matrix $T$ which has a peculiar quasi-Hankel structure. For example, \begin{equation} T_2= \left( \begin{array}{c|ccc|cccccc} a & b & c & d & e & f & g & h & i & j\\\hline b & e & f & h & 0 & 0 & 0 & 0 & 0 & 0\\ c & f & g & i & 0 & 0 & 0 & 0 & 0 & 0\\ d & h & i & j & 0 & 0 & 0 & 0 & 0 & 0\\\hline e & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ f & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ g & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ h & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ i & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ j & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{array} \right). \end{equation} \begin{equation*} T_3=\left( \begin{array}{c|ccc|cccccc|cccccccccc} a & b & c & d & e & f & g & h & i & j & k & l & m & n & o & p & q & r & s & t\\\hline b & e & f & h & k & l & m & o & p & r\\ c & f & g & i & l & m & n & p & q & s\\ d & h & i & j & o & p & q & r & s & t\\\hline e & k & l & o\\ f & l & m & p\\ g & m & n & q \\ h & o & p & r\\ i & p & q & s\\ j & r & s & t\\\hline k & \\ l & \\ m & \\ n & \\ o & \\ p & \\ q & \\ r & \\ s & \\ t \end{array} \right) \end{equation*} Larger matrices $(T_4,T_5,...)$ have the same block structure; in fact, each successive $T_i$ is contained as a partial block of all $T_j,j>i$ (see above examples). All of the unique elements are contained in the first column or row, but the matrix does not possess classical Hankel/Toeplitz/etc. structure typical of fast structured matrix vector multiplication. This matrix arises from a convolution of a certain sort, so I am convinced that the matrix-vector product can be computed in $\mathcal{O}(N\log N)$ time or something close rather than $\mathcal{O}(N^2)$. I'd appreciate the input of others, or potentially helpful references.

Edit: The block structure is controlled by the parameter $P$ so that the matrix $T_P$ has a $(P+1)\times(P+1)$ upper left triangular block structure. The $p$th row or column block is of dimension $(p+1)(p+2)/2$.

$\endgroup$
  • $\begingroup$ The product can be performed in almost linear time. Look this reference. $\endgroup$ – The Doctor Dec 15 '17 at 22:36
  • $\begingroup$ I am aware of those types of algorithms but the standard algorithms referred to in the reference do not apply here due to the particular structure of these matrices. I think maybe my use of "quasi-Hankel" might be confusing. All I meant is that the blocks along the upper block antidiagonals contain the same matrix elements, not necessarily that the matrices themselves are identical. $\endgroup$ – sssssssssssss Dec 15 '17 at 23:09
  • 1
    $\begingroup$ Looking at $T_3$, doesn't it mean your matrix has $<N\times (P+1)$ nonzero elements, $\ll N^2$? So the direct straightforward matrix-vector product will take $N(P+1)$ operations. I'm not sure how $P+1$ compares to $\log N$, but it's not clear to me that you even need any complicated algorithm here. $\endgroup$ – Kirill Dec 16 '17 at 17:41
  • $\begingroup$ The complexity is certainly $P^6$ when worked out for $T_P$. The constant is small but the scaling is suboptimal $\endgroup$ – sssssssssssss Dec 29 '17 at 1:32
4
$\begingroup$

Late answer but the matrix can be thought of the following sum of rank-$2$ matrices which allows for having smaller block sizes as the number of blocks increase.

$$ \begin{align*} c_{20}=\left( \begin{array}{cc} \frac{a}2 & 1 \\ b & 0\\ c & 0\\ d & 0\\ e & 0\\ \vdots&\vdots\\ r & 0\\ s & 0\\ t & 0 \end{array} \right) \left( \begin{array}{cccccccccccccccc} 1&0&0&0&0&\cdots&0&0&0&\\\frac{a}2 & b & c & d & e &\cdots & r & s & t \end{array} \right)\begin{pmatrix}x_1\\\vdots\\x_{20}\end{pmatrix}+\\ \left( \begin{array}{cc} \frac{e}2 & 1 \\ f & 0\\ h & 0\\ k & 0\\ l & 0\\ m & 0\\ o & 0\\ p & 0\\ r & 0 \end{array} \right) \left( \begin{array}{cccccccccccccccc} 1&0&0&0&0&0&0&0&0&\\\frac{e}2 & f & h & k & l &m & o & p & r \end{array} \right)\begin{pmatrix}x_2\\\vdots\\x_{10}\end{pmatrix}+\\\textrm{etc.} \\ \end{align*} $$

Note that, by performing the second multiplication at each term, most of the calculations are vector times scalar products and "get the first element" type operations hence allows for significant reduction. Thus this has one inner product of decreasing size, two lookups and one vector times scalar and two scalar addition to an element operations.

I suspect this matrix could have been formulated as an iterative solution instead of the linear set.

$\endgroup$

Your Answer

By clicking "Post Your Answer", you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.